AI in professional services: the input is a billable hour
Law firms, accountancies and consultancies have a constraint no other sector on this site shares: the knowledge a system needs is held by the people whose time the firm sells. In a bank or a hospital, the domain expert who defines what a correct answer looks like is on salary, and an hour of their attention costs the project nothing visible. In a firm that bills by the hour, that same hour has a price attached, appears on a utilisation report, and competes directly with revenue. This makes evaluation work, which is the foundation of everything else, structurally harder to fund than the build. The second constraint follows from the same fact. The obvious use case is drafting, it works well, and it makes the firm's people faster at precisely the activity the firm charges for, which is a question for a partner rather than for a project team.
Professional services, in short
| The constraint | The knowledge being automated lives in senior people who are billable, which makes the input a system needs the most expensive resource in the firm. |
|---|---|
| The distraction | Drafting, which works well and produces value that is hard to capture when the business model bills for time. |
| A first project that works | Retrieval across the firm’s own prior work, which is a search problem the firm has failed at for twenty years. |
| Where ground truth lives | Prior matters and their outcomes, which exist in a document management system nobody has indexed properly. |
What shows up most often, not a description of any particular organisation. No named clients and no case studies: see editorial policy.
The billing question has to be answered before the build
A firm that bills for time and deploys a drafting system has created an unresolved tension, and pretending otherwise does not make it go away. If a task that took four hours now takes one, either the client is billed for one hour or the firm has a conversation it has been avoiding.
Firms resolve this in different ways and all of them are decisions somebody senior has to take: moving the work to a fixed fee, using the recovered time for work that was being skipped, or accepting lower recovery on that task in exchange for winning more of it. Any of these is workable. None of them is a project team's decision.
The failure mode is a successful pilot that quietly stops, with no stated reason, some months after a partner looks at the numbers. Teams experience this as inexplicable and it is entirely predictable, because the question was never put to anyone empowered to answer it.
Retrieval across prior work sidesteps all of it
Almost every firm has been failing at this for twenty years. The knowledge is there, in past matters, advice given, precedents and engagement documents, and finding it depends on someone remembering that a similar question came up in 2019.
This is a problem current technology handles genuinely well, and it has three properties that make it the right first project here. It does not touch the billing model, because it makes people more effective rather than faster at a billed task. Its value is obvious to anyone who has ever hunted for a precedent. And the ground truth exists: prior matters and their outcomes are in the document management system already.
It also produces a political benefit that matters in partnerships. A junior who finds the right precedent in a minute rather than a morning is visibly better at their job, and the person whose prior work was surfaced gets the credit, which makes the system something people want rather than something imposed.
Information barriers are a permission problem, not a filter
Conflicts of interest and client confidentiality create real obligations about who may see what, and they are enforced by professional regulators rather than by policy.
A retrieval system that indexes everything and filters results afterwards will eventually leak, and it will leak in the worst possible way: a user seeing a fragment of a matter they are walled off from. The design that holds is one where the retrieval itself runs under the user's own permissions, so unavailable material is never a candidate.
This is a day-one architecture decision. Retrofitting it is close to a rebuild, and in this sector the consequence of getting it wrong is not an embarrassment but a regulatory matter.
Buying the expert hour deliberately
Since senior time is the scarce input, it should be budgeted like one. Naming the partner or senior associate who will own correctness, agreeing how many hours that is, and getting it approved before the project starts is the single practice that separates firm projects that measure anything from ones that argue about quality.
The efficient version is narrow. Twenty real examples argued over for two hours produces a better evaluation set than a specification meeting, and it is a cost a firm can see and approve. The five stages hold otherwise, with stage one unusually cheap here because the experts are articulate and stage three unusually dependent on their availability.
